|
Joseph Mack wrote:
> > If not, I'd like to hear what other
> > folks are using for benching LVS.
>
> I'm in the middle of trying to figure all this out. The short answer
> is we don't have anything real good. Conventional clients make one request
> at a time. To get multiple requests you need multiple client processes, or
> client programs being run on multiple hosts. You then have the problem
> of merging the logs from the clients. As well the client logs are usually
> not machine readable. Few are ready for gnuplot.
Polygraph is what you need. When setup properly it can simulate
hundreds of clients and servers from a single client-server process pair
(which should be run on separate machines but for low loads of <100
reqs/sec I've found it works fine running both on a single machine).
There are some very good log extraction and analysis tools included in
the src distribution. It's not immediately user friendly but once you
find your way around it's very powerful, and it is Gnuplottable using
the included Report_Generator.
All of your problems are answered pretty well by polygraph (including
merging and analyzing multiple client/server logs so you can test
massive loads...like thousands of requests per second).
Polygraph has its flaws: First, forget running workloads that include
"pipe" features, which allows simulation of slow and lossy lines--like
56k modems--on a Linux box...Only FreeBSD seems to work with this.
Second, creating a workload is a complex and tedious task that is really
only understood by Alex! Finally...I don't know how far along the L4
Switch workload is towards completion, but I must assume it is at least
available in a development version.
Anyway, I've found it to be priceless in my quest to squeeze blood from
my Squid servers. ;-) And it has helped uncover several bugs that had
never shown their face under mere mortal loads.
Until there is a large effort underway to benchmark and optimize LVS,
I'm betting there are a lot of areas where things are not as fast or
efficient as they could be. (Squid is a prime example...development has
speeded up markedly since the cache bakeoffs began, and very big
improvements are being made to stability and speed.)
I'll be happy to volunteer bench time here in the lab for testing LVS as
soon as there is a good workload to test it with. (We have two
polygraph stations running 24 hours a day, no reason why some of that
time can't be spent on something other than Squid.)
Talk at you later.
--
Joe Cooper <joe@xxxxxxxxxxxxx>
Affordable Web Caching Proxy Appliances
http://www.swelltech.com
|
